Research On Instance-based Inductive Transfer Learning

In the tide of global informationization,under the impetus of the era of big data,the growing amount of data,its growth is impressive,with so much information and data of sudden impact on the brain,people don't have time to see these data,their attention has turned to a valuable means to cope with.As a result,people have to find a effective method to deal with a lot of information to efficient and accurate to find valuable information.First of all,to the relevant data,this paper consider the hierarchy relationship between data,adjustme similarity between the hierarchy nt weights of data correlation with the target field on the basis of correlation by different proportion.Import similarity between the hierarchy,weight adjustment strategy,wrong sample weight constraints in the classic transfer learning algorithm,TrAdaBoost.So as to put forward a transfer learning algorithm combined with the similarity between the hierarchy in order to solve a hierarchy relationship data and how to efficiently adjust the weight problem.In the experimental part,this paper will prepare accuracy,precision and recall rate between transfer learning based on hierarchy correlation ? SVM,TrAdaboost algorithm,analysis of experimental data based on hierarchy correlation data have more accurate classification results.Aimed at the similarity but not same between the target domain and the multiple source domain,the multiple source is also different,they are multiple source in a large domain with some relevance.Aimed at the problems of multiple areas,the amout of samples in some single source domain is not enough,negative transfer,the paper rise a multiple source instance transfer learning algorithm.The algorithm consider knowledge from many source domain to make the target domain can consider to use the knowledge of every source domain.The method firstly combine source domain and target domain to train a classifier and save the source domain which improve classify effect after test.And then,we train the leaving source domain assemed used by Tradaboost.After the test,we choose the final set as the sourcedomain according to specified rules to train classifier with target domain together.Finally,we describe the hierarchy relevant transfer learning algorithm in third chapter and multi-source instance transfer learning in fourth chapter we have raised in detail through the experiment.Then we make objective analysis summary to experimental results.We compared before and after improvement of accuracy,precision and recall.Proved that the hierarchy relevant transfer learning algorithm and multi-source instance transfer learning algorithm can effectively improve the classify effect of classifier.